Research On Semantic Segmentation Based On Improved PSPNet

Semantic segmentation covers many scientific fields such as image processing,deep learning,computer vision and human-computer interaction,has great economic value and wide application prospects.With the rapid development of artificial intelligence and deep learning,semantic segmentation has also made breakthrough progress.In the early stage of semantic segmentation,the best semantic segmentation model is Fully Convolutional Network(FCN).However,in the experiment of ADE20 K data set,FCN divided the ship into cars,resulting in the mistake of "cars on water".Therefore,Pyramid Scene Parsing Network(PSPNet)is based on the idea of FCN,introducing more context information,so that pixels in the image no longer exist in isolation,and there are more semantic information between adjacent pixels,and the probability of false segmentation will be significantly reduced.Different from the traditional PSPNet model,we improve the network structure of PSPNet and introduce the traditional segmentation method as post-processing.Firstly,in feature extraction,the traditional Res Net model relying on convolution operation is replaced by Mobile Net V2 model relying on depth separable convolution,which greatly reduces the amount of parameters.At the same time,multiple feature layers of Mobile Net V2 model are screened.With the help of attention mechanism,four feature maps which have the greatest influence on the segmentation results in the global receptive field are selected for extraction,expansion and fusion,so as to retain more learnable features.After that,the obtained feature map is divided into four sub-regions of 1×1,2×2,3×3 and 6×6,which are averaged and pooled respectively.The length and width of the obtained four groups of results are uniformly defined,and the network results are fused and output again.In the post-processing part,we obtain the initial contour of the level set method through binarization,filling,contour extraction and other operations.Through the iteration of the level set method,the border is closer to the target edge,improving the accuracy,and obtaining the final result.In this paper,we adopted the Baidu People Segmentation dataset for human body segmentation.The dataset consists of 5387 training pictures.During training,600 pictures are randomly selected from 5387 as the verification set,and then 4787 pictures are used as the training set for training.And Adobe's Portrait Segmentation dataset,which consists of 1500 training sets and 300 test sets.Experiments show that the pixel accuracy of the proposed method reaches 94.47% and 93.01%,respectively,which has a significant improvement effect on the accuracy.The average pixel accuracy is increased by 5%,which has high research value.